Obviously, the measure isn’t mistakes per day, it’s mistakes per LOC. And that’s not the whole story either - AI self-corrects in addition to being corrected by the operator. If the operator’s committed bugs/LOC rate is as low as the unaugmented programmer’s bugs/LOC, you always choose the AI operator. If it’s higher, it might still be viable to choose them if you care about velocity more than correctness. I’m a slow, methodical programmer myself, but it’s not clear to me that I have a moat.